TECH

'AI is incompatible with democracy,' says author of book on OpenAI

One aspect that has surrounded the field of artificial intelligence (AI) since the ChatGPT boom is the aura of "magical technology"—something that emerged almost spontaneously and carries promises and concerns as grand and inevitable as anything science fiction could imagine.

That is not how American journalist Karen Hao views the field. To her, it is a project with the hallmarks of an empire, consuming global resources on a massive scale to benefit a tiny elite in Silicon Valley, USA. Her research led her to conclude that AI is incompatible with democracy—a position she argued in a book that garnered significant attention last year.

She is the author of *The AI ​​Empire* (originally titled *The AI ​​Empire* in English), which details the history of the company co-founded by its current CEO, Sam Altman. A Portuguese-language edition of the book was recently released.

Her perspective stems from a unique vantage point: as the former AI editor for *MIT Technology Review* and a reporter for *The Wall Street Journal*, Hao closely followed the emergence and rise of OpenAI. In just a few years, the organization transformed from a non-profit laboratory into a company currently pursuing a $1 trillion IPO.

The book maps out the key figures, companies, technologies, dilemmas, and issues in the field. It serves as a guide to understanding how a technology that once seemed like the stuff of movies became a major driver of political, economic, cultural, and behavioral influence in recent decades.

Hao is currently in Brazil. On Tuesday (the 30th), she is participating in an event leading up to the 21st International Investigative Journalism Congress, organized by the Brazilian Association of Investigative Journalism (Abraji). The talk begins at 4:00 PM at the Camargo Guarnieri Cultural Center on the University of São Paulo (USP) campus. The event is hosted by Abraji, Editora Rocco, and GEIA (the Research Group on AI and Digital Cultures) from USP’s School of Communications and Arts. Yesterday, she met with *O Globo* at a hotel in São Paulo’s West Zone and, in addition to discussing her book, addressed some of the latest topics in the world of AI. These included data center regulations, OpenAI’s market position, Anthropic’s growing strength, the US government’s ban on advanced models, the papal encyclical, the lack of diversity in AI research, and the global influence wielded by Silicon Valley companies.

https://karendhao.com/

Read the highlights of the conversation below.

Throughout the book, Sam Altman and those around him speak as if everything regarding AI is inevitable, yet you counter that view. Is there anything about AI that is truly inevitable?

I don’t think anything is inevitable in general, but when it comes to AI, one of the things I really tried to highlight in the book is the extent to which every decision regarding ChatGPT was based on completely subjective choices. And it’s interesting when you look at OpenAI’s beginnings.

The approach they took to scaling their LLMs was seen as scientifically anomalous at the time. They adopted a brute-force, intellectually "cheap" approach that wasn't the path favored by other researchers in the field. And it was partly because they had an extraordinary amount of money that they managed to make their approach the dominant one. So, how can you say that a technology born from that kind of history is somehow inevitable?

In what ways does AI undermine democracy?

I call these companies—like OpenAI—"AI empires" because of the striking parallels they share with the empires of the past and the way they amass extraordinary economic and political power.

This happens through the dispossession of the majority. They dispossess people of their data, their land—to host these data centers—, water resources—to power and cool these centers—, as well as people's labor, future economic opportunities, and educational opportunities. And that is why the empire manages to extract an extraordinary amount of value so quickly: it extracts it without distributing it back proportionally. And the reason I believe this threatens democracy is that empires and democracy are incompatible.

Empire is founded on the idea that there is a natural hierarchy in the world—that there are superior and inferior groups—and that those at the top deserve to be there and deserve to appropriate all those resources by virtue of some divine right or natural order. Democracy, on the other hand, is based on the exact opposite premise: the idea that we are all equal and all deserve to participate collectively in determining our own future. Thus, at a purely philosophical level, there is a fundamental conflict between the ideology driving the development of the AI ​​industry and the way democratic societies are organized.

Brazil is working on legislation to attract data centers. By offering cheap renewable energy and tax incentives, the country is positioning itself as a “digital colony,” as you describe in your book. What can Brazil learn from countries like Chile, which have both welcomed and rejected data centers?

One notable trend we’ve seen over the past year is the rise of resistance to data centers worldwide. It started in Latin America—in places like Chile—and spread to the US, Europe, and Brazil. This kind of grassroots organizing is really beginning to pressure the AI ​​industry to change its approach.

For instance, OpenAI shelved Sora (its AI video generator). When they announced the product, they billed it as the most important launch since ChatGPT, yet within a few months, they had to shelve it. Grassroots organizing was the reason why. If you look at the three reasons reported for OpenAI’s decision, the first was a massive bottleneck in computing power.

The second reason was stagnant consumer demand—so, this is a case of collective consumer action. The third is that OpenAI is preparing for an IPO and facing a much more uncertain financial landscape. Wall Street is increasingly skeptical about whether the AI ​​industry can actually deliver on its promises, given the massive political and social backlash currently underway.

This is where Brazil can gain insight: by recognizing that when this kind of grassroots organizing and resistance emerges—whether against the infrastructure, the way they harvest data or intellectual property, or the psychological harm inflicted on children—it impacts the trajectory of AI development.

The US government is making it very clear that it decides who gets access to cutting-edge technology—as seen with Claude Fable. What should countries do to preserve their sovereignty while still keeping pace with the latest advancements?

There is a major question here regarding why we actually want to keep up with the latest technologies. If these latest technologies are the same ones exploiting and extracting resources from communities around the world, is it really a good thing to keep up with them? Or should we, in fact, reframe the problem regarding the rules of the game? If we were to redefine our goals—not just to chase the latest tech, but to pursue the objectives of individual communities, such as improving the cost of living, the quality of education and healthcare, and economic opportunities—you would quickly realize that we don't need any of the AI ​​technologies Silicon Valley is trying to force down people's throats.

There is a completely different set of AI technologies we should actually be developing. And there are many ways to develop these technologies without engaging in Silicon Valley's exploitative practices. This would help communities continue to progress in the true sense of the word—not just technological progress for its own sake, but human, social, and economic progress.

Right now, OpenAI is squeezed between Anthropic—which currently has the most popular platform—and SpaceX, which potentially has the capacity to build its own infrastructure. These are two things I don't see happening for OpenAI. Will the empire fall?

If we define the "AI empire" solely as OpenAI, then yes. There is a lot of pressure on OpenAI right now, and it doesn't seem to be in a very strong position. But I think the more important question is: will the AI ​​empires—plural—fall? I am actually quite hopeful about this, because I define OpenAI, Anthropic, SpaceX, Amazon, Microsoft, Google, and Meta all as AI empires. Having OpenAI cease to be an empire only for Anthropic to take its place wouldn't actually solve the core problem I see regarding the destructive and exploitative nature of AI development. What I hope happens is not merely that we keep swapping which entity acts as the dominant empire, but rather that—through grassroots organizing and resistance—we secure genuine accountability from all these empires. I want them to stop being empires and instead become companies that offer value commensurate with what they receive in return. My goal is not to put these companies out of business; it is simply to bring them back to a role where they are not excessively exploitative and do not degrade the environment. We can have companies that provide high-quality products and services without causing extraordinary amounts of damage.

This approach is incredibly circular and baffling, because when you look at what Anthropic is doing, it’s practically the same thing as OpenAI. Dario Amodei and the Anthropic executives left OpenAI, essentially copied and pasted their model, and simply rebranded themselves as the "good guys" without actually addressing the root causes of the problems. Both OpenAI and Anthropic approach AI through scaling, so you still run into the same issues regarding data privacy violations, the erosion of intellectual property, environmental damage, and harm to public health. They are essentially making minor tweaks and claiming moral superiority, when in reality, they are just another empire.

What do you make of the fact that the Pope relied on someone from Anthropic to help present his encyclical, *Magnifica Humanitas*? Leo XIV warned about the threat AI poses to human dignity, justice, and labor, yet he had a representative from one of these companies by his side.

That was a very confusing moment for me, because I found the encyclical to be an incredibly profound and beautifully written document. It discusses how AI represents a new phase of colonialism and perpetuates labor exploitation—potentially giving rise to new forms of slavery—and addresses how the AI ​​industry advances based on an ideology that assumes machines will always be superior to humans and will somehow perfect the flaws of the human species. One of my favorite lines says, "We flourish within our limitations, not despite them." Yet there was Chris Olah, an Anthropic executive, standing right alongside the Pope.

At first, I was quite disappointed and thought, "Okay, so even the Vatican—the Catholic Church—has, in a way, surrendered to these companies." But when Olah spoke, he framed the document as a critique. He didn't say, "We fully support this document." In fact, he said, "We need critics like the Pope to hold companies like Anthropic accountable." So, he framed it as an adversarial dynamic.

On one hand, Anthropic is trying to draw the Vatican closer to its sphere of power and influence, but at the same time, the Vatican is trying to do the same by keeping the AI ​​industry in check. It isn't entirely clear who actually gained more from this arrangement. Was it the AI ​​industry, or was it the Pope? In the end, perhaps the Pope managed to have the final word.

Is Artificial General Intelligence (AGI) a lie? Why isn't it discussed with the same intensity anymore?

It is a myth, in the sense that it is an incredibly compelling story that many people believe in—and one that serves a huge political purpose for these companies to justify all the destruction they cause. If they can get everyone to believe that AGI somehow exists, then they can simply go on doing whatever they want. But reality has set in, and myths only work in information vacuums. The more AI is deployed in society, and the more communities suffer the impact of its development, the more that information vacuum gets filled with actual facts about what the technology really represents—and the more that myth begins to crumble.

You argue that the scientific field of AI has lost transparency and research diversity with the rise of generative AI. How can that be regained?

The reason so much diversity has collapsed is that the AI ​​industry has become the dominant funder, even outside of the companies themselves. They are the primary funders of academic labs, while state funding accounts for a smaller share of the total.

Changing this requires a few things. One is having people with alternative visions for AI development who refuse to accept funding from these companies. New sources of funding will also be required, potentially combining state funding, foundation support, and perhaps even crowdfunding. A richer constellation of startups will also be needed. This will take time.

Restoring the diversity that was lost will require a great deal of careful effort and sacrifice, as AI researchers and other talented individuals will need to make an active choice early in their careers: forgo a $1 million compensation package to invest in a different approach to AI development.

Is there a specific area of ​​AI where you would like to see more research—beyond deep learning, machine learning, and generative AI?

Before deep learning became the primary focus of virtually all AI development, there was a field known as neuro-symbolic AI. This approach involved encoding knowledge and databases into computer systems to create a more deterministic system—one that could retrieve that knowledge and reason through the database to arrive at specific answers. While that approach had its weaknesses and was eventually sidelined due to being too slow and costly, the neuro-symbolic school of thought merges the strengths of deep learning with those of the symbolic approach. It allows a system to learn quickly from data—as deep learning does—while also incorporating fundamental rules that do not need to be learned. We already know that 1 plus 1 equals 2, so there is no need to feed the system vast amounts of data demonstrating this fact. That is part of the reason why deep learning systems are so inefficient; they essentially reinvent the wheel every time. So, I am interested in seeing more work on neuro-symbolic AI.

At the same time, within the realm of deep learning, there were other interesting avenues for making systems more efficient—reducing their consumption of data and computing power. I believe there is significant research to be done both within and outside the current paradigm to explore the new techniques and methodologies we can use to achieve the desired capabilities without relying on an extractive supply chain.

Finally, I would add that the issue isn't just how we achieve better systems, but also how we define the ultimate goal. For some time now, the industry has defined that goal as replicating human intelligence. I don’t think that’s the right goal. The aim of technological development is to complement what we cannot do, not to replace what we can do. When we got the first computers and calculators, part of why that was great was that humans can’t calculate numbers as fast as computers can. So, we offloaded that work to the computer, but there are so many other things humans can do that we would never be able to offload to a machine. Why not focus solely on developing AI systems geared toward the things we could never do ourselves, rather than trying to outdo us in every way?

Do you use AI tools? Which ones, and for what purpose?

I don’t use commercial generative AI tools. I don’t use ChatGPT, Claude, Gemini—none of them—and there are three reasons why. First, because I investigate these companies, so from an ethical standpoint, I don’t want to participate in perpetuating the harmful practices they engage in. Second, for privacy reasons. I investigate these companies, so I don’t want to hand over all my data to them. And third, because I believe that, ultimately, the strengths of my work are simply incompatible with what I would get from a generative AI tool.

However, I do use specialized AI tools. For example, one of the things I wanted to do with my book was detail how OpenAI became better capitalized after shifting from a non-profit organization to a Microsoft-funded venture, and I noticed a huge improvement in their office furnishings.

The office chairs I saw at the first location were simply very different from the ones at the next office. So, I took photos of each chair and ran them through Google Image Search—a specialized AI tool that doesn't try to generate anything and doesn't consume vast resources to perform the task. I discovered that the chairs from the old office actually cost $2,000 each, while the chairs in the newer office were by a famous Brazilian designer and cost $10,000 apiece. I included this in the book, as I felt it helped illustrate the point.

--o Globo--